Cross-subject emotion recognition in brain-computer interface based on frequency band attention graph convolutional...

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Cross-subject emotion recognition using electroencephalogram (EEG) data is challenging due to inter-subject physiological variations.
• The 'domain gap' in EEG data distributions across subjects is a key obstacle for accurate emotion recognition.
• Addressing this domain gap is crucial for developing robust and generalizable emotion recognition systems.

Purpose of the Study:

• To develop a novel model that effectively narrows the domain gap in EEG data for cross-subject emotion recognition.
• To leverage emotional frequency bands and inter-channel relationship information to improve model generalization.
• To enhance the performance of emotion recognition systems across diverse subject populations.

Main Methods:

• Proposed the Frequency Band Attention Graph convolutional Adversarial neural Network (FBAGAN) model.
• Employed a feature extractor with a frequency band attention mechanism and a graph convolutional neural network (GCN) to capture spectral and spatial information.
• Utilized a discriminator to minimize the distribution discrepancies between source and target domains in terms of frequency and channel relationships.

Main Results:

• FBAGAN achieved high accuracy on benchmark datasets: 88.17% on SEED and 77.35% on SEED-IV.
• The model demonstrated strong performance on DEAP dataset for Arousal (69.64%) and Valence (65.18%).
• Results indicate superior performance compared to most existing cross-subject emotion recognition models.

Conclusions:

• FBAGAN effectively bridges the domain gap in EEG channel and frequency band distributions.
• The proposed model demonstrates significant improvements in cross-subject emotion recognition performance.
• The findings highlight the potential of attention mechanisms and GCNs in addressing domain shift challenges in EEG analysis.